Research map / Vision-language and multimodal / Multimodal LLMs

HoneyBee: Data Recipes for Vision-Language Reasoners

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematically studies VL reasoning CoT data curation (sources, interventions, scaling) with fixed SFT and evaluation, and releases the 2.5M-example HoneyBee dataset that trains PLM-1B/3B/8B VL reasoners.

Problem

Open questions remain about how to construct high-quality vision-language chain-of-thought reasoning datasets, since state-of-the-art VLM data recipes are proprietary and prior comparisons confound data, generator, and training choices.

Approach

Fixes the training algorithm, CoT generator (Llama-4-Scout), and evaluation suite to isolate the effect of context sources, data interventions (caption-and-solve, text-only mixing, filtering, etc.), and scaling along images, questions, and CoTs; combines the winning choices into a 2.5M-example dataset.

Key results

• PLM-HoneyBee-1B reaches 36.2 average vs 28.3 for InternVL3-1B-Instruct across 10 benchmarks.
• Caption-and-solve augmentation improves 5-benchmark average by 3.3pp; adding text-only reasoning data improves it by 7.5pp.
• Training on the VirL source alone outperforms other context sources by up to 11.4pp average gap; mixing sources did not help.

Surprising findings

• Mixing data from diverse VL sources did not improve over the best single source (VirL).
• Most data interventions (visual perturbation, text-rich images, perceptual redundancy filtering, increased distractors, length, difficulty) failed to beat the original baseline data.

Limitations

• Experiments restricted to single-image math-centric reasoning data.
• Data curation decisions rely on a limited set of validation benchmarks (five) used for hill-climbing.
• Single generator (Llama-4-Scout) fixed throughout, so findings may depend on teacher choice.

Experimental setup

Datasets: VirL-39K MathV360K R1-Onevision ThinkLite LLaVA-CoT MMK OpenThoughts-1M text reasoning

Benchmarks: MathVerse MathVista MathVision MMMU-Pro We-Math DynaMath LogicVista HallusionBench MATH500 GPQA

Metrics: accuracy

Baselines: InternVL3-1B-Instruct InternVL3-8B-Instruct Qwen2.5-VL-3B-Instruct Qwen2.5-VL-7B-Instruct OpenVLThinker-v1.2-3B LLaVA-CoT MMK ThinkLite R1-Onevision MAVIS-style VirL

Builds on: PLM (Perception Language Models) GRPO STaR OpenThoughts self-consistency

Closest papers by meaning

• Prism: A Framework for Decoupling and Assessing the Capabilities of VLMs (NeurIPS 2024 · similarity 0.90)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.90)
• Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale (ICML 2026 · similarity 0.90)
• Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs? (ICML 2025 · similarity 0.90)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
• Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning (NeurIPS 2024 · similarity 0.89)